Executive Summary

Key Figures

Source: LAION-BVD paper (arXiv 2608.24845), body text and Table 1

LAION-BVD begins not with video but with addresses. The team pulled 4.7 billion candidate video links out of every CommonCrawl dump available as of March 2024, then narrowed the pool to three platforms, YouTube, Vimeo and Dailymotion, leaving 1.3 billion URLs. The paper gives two reasons for keeping only those three. One is reliable access. The other is that these platforms moderate their own harmful and unsafe content.

Then came the collection itself. Some 2,000 virtual servers were coordinated through a Celery cluster and fed by yt-dlp. To keep access from being throttled, the pipeline ran through a residential proxy network. Roughly 130 million downloads were attempted, about 60 percent succeeded, and the result is 80 million videos totalling 10 million hours.

What that material looks like is nothing like the tidy research corpus most people picture. Of the haul, 94 percent came from YouTube, and the largest content categories are People & Blogs at 20 percent and Music at 17 percent. English accounts for 57 percent of the audio, with Russian at 9 percent and Spanish at 8 percent, so close to half is non-English; Korean makes up 2 percent. The mean video runs 7.7 minutes and the median 3.7, and upload dates spread across nearly two decades, from 2006 to 2024. This is not a set of curated experimental clips. It is the web's video as it sits.

The first gap in the record opens right here. The paper states plainly that no additional safety filters were applied at the collection stage, because the source platforms already moderate for harmful and unsafe content. Filtering was delegated wholesale to someone else's pipeline. LAION has its own history here: when suspected child sexual abuse material was found among the links in LAION-5B in December 2023, the dataset was withdrawn, and a cleaned edition (Re-LAION-5B) followed eight months later. Against that history, the filter this time has not been made finer. Its place has moved out to the platforms.

Downloaded video does not go into training as is. PySceneDetect finds the cuts and slices the footage into clips, motion is estimated on low-resolution frames, and clips that turn out to be effectively static are thrown away. From a random sample of 2.4 million videos the pipeline produced 55 million clips. Each clip then needed a description, and three models wrote them.

Every modality got its own captioner and its own instruction. What follows is the whole of the captioning pipeline as Section 3.2 of the paper describes it.

Qwen3-VL-2B, the video captioner, is a small model at 2 billion parameters. The paper's reason for not reaching for something larger is scalability: 55 million clips have to be worked through, so the model has to be light. The twenty-word ceiling follows the same logic, and the ten-word limit on audio captions was chosen to match the average caption length in existing audio-text datasets such as AudioCaps and Clotho, around 8 and 11 words respectively.

Nor was the captioner itself picked by human reading. For the frame captioning model the team started from the top of the OpenVLM leaderboard and, held to what its compute and GPU memory allowed, cut the field to seven candidates under 4 billion parameters. Each one captioned 100,000 images from DataComp-1B, and those captions were scored against the images with CLIPScore. DeepSeek-VL2-tiny took the job on the highest score, 0.62, and the second-best throughput. The yardstick for choosing which model would write the descriptions was, in other words, another model's score.

2.1None of This Is New

LAION-BVD did not invent the practice of letting a model write the labels. Of the 19 open video-text datasets compared in the paper's Table 1, exactly two list a human caption source: MSR-VTT and DideMo. The rest are already model-generated, or speech-recognition transcripts, or alt-text. Human-written labels became the exception somewhere around InternVid in 2023.

What changed is the surface area. InternVid, the previous largest, held 7.1 million videos and 760,000 hours; LAION-BVD holds 80 million videos and 10 million hours. The same convention now runs across ground more than ten times larger. There is now more video carrying a description no human has ever read than there is in every open corpus that came before it, combined.

So who checked whether those descriptions match the footage? LAION-BVD never looked at the captions directly. It trained models on them and measured the scores. ViCLIP handled video-text, CLAP handled audio-text, CLIP handled frame-level image-text, and each went onto the standard benchmarks.

The numbers are respectable. Holding the sample budget at 10 million, a ViCLIP trained on InternVid-10M-FLT reached an overall average of 58.0, while the BVD-V-10M model reached 61.3. That average pools zero-shot classification on K400, UCF-101 and HMDB51 with retrieval on MSR-VTT and MSVD. The comparison comes with a condition attached, though. InternVid's originally reported numbers used checkpoint merging (WiSE-FT), and applying it to both sides gives 60.2 against 62.3, narrowing the gap to 2.1 points.

At the frame level the measurement was distributional rather than a score. The distance (FID) between LAION-BVD frames and Re-LAION images measured 33.92, whereas two samples drawn from within Re-LAION itself sat 0.16 apart. Frames pulled out of video are statistically a different picture from the web images we already had.

The audio results deserve a more careful reading. BVD-A-10M, trained on synthetic audio captions alone, matched or beat the existing LAION-Audio at all four model sizes. But the combination that mixes in AudioSet, curated specifically for audio tasks, scored 55.5 to 56.6 under the same conditions, while the synthetic-caption-only runs stayed between 42.9 and 46.8. A gap of roughly ten points that did not close as the models grew. Scale did not catch up with curation so much as fill the space where curated data was absent.

The question is what this validation actually answered. A higher benchmark score means models trained on this data perform better. It does not mean the captions describe the video accurately. The two propositions overlap broadly but they are not the same. A systematically biased label will still lift the score whenever its bias points the same way as the evaluation benchmark. The paper does not hide the gap.

As a research disclosure that is honest. Move the same sentence into an organization's data quality standard, however, and it changes character. Human-labelled datasets come with annotator agreement rates, re-review percentages, distributions of error by type. Synthetic labels have no standardized equivalent yet, and downstream performance has moved into the vacancy. Downstream performance is an outcome measure, so it cannot tell you where or how a label went wrong.

The appendix does contain one experiment aimed squarely at that gap. The images were held constant and only the captions were swapped. A CLIP trained on DataComp-1B with its original alt-text was compared against a CLIP trained on the same images after this pipeline's captioner had described them afresh. At 128 million samples, zero-shot accuracy on ImageNet-1k fell from 0.40 to 0.23, while COCO retrieval recall rose from 0.33 to 0.42. Same photographs, same training recipe, and only the author of the description had changed.

The paper also counted where that difference comes from. Across a 300-million-sample subset of synthetic captions and a matching subset of DataComp alt-text, ImageNet-1k class names appeared roughly 21 million times in the former and 141 million times in the latter. A captioning model narrates the scene smoothly and rarely names the object; the alt-text a person typed into a web page is short and rough but calls things by their names far more often. That is how concretely the identity of the label's author changes the character of a dataset. Synthetic labels leave a trace, and the trace can be counted.

"Open dataset" means two different things here. What went up on HuggingFace is the video URLs and a portion of the captions. The footage itself is another matter. It reaches only verified researchers who accept the terms of use and work inside a closed research data environment.

Read those terms and the governing provision appears in the opening paragraph: Sec. 60d(4) of the German Copyright Act (UrhG). It is Germany's national implementation of the text and data mining exception for scientific research in Article 3 of the EU's Digital Single Market copyright directive, and it covers making material available to a specifically delimited group of researchers, inside a closed environment, for a particular joint research project. Scientific research, as the definitions section pins down, means non-commercial research.

Hence the run of clauses sealing off commercial use. A few of them read as follows.

• Any directly or indirectly commercial use is prohibited, and that expressly includes training, validating or fine-tuning AI models for commercial purposes (Sec. 5).
• Redistribution and third-party access, re-identifying or profiling the people shown in the videos, and training prohibited biometric identification or surveillance systems are all barred (Sec. 6).
• When the research ends, copies of the dataset must be deleted, and the deletion confirmed upon request in writing or in text form (Sec. 11).

That last clause carries an exception practitioners should note. Lawfully created derived results and models, in particular trained models, model weights and embeddings, fall outside those deletion, purpose and time restrictions. The raw material goes back but whatever was built from it stays. For anyone whose job is tracing data lineage, the point where the deletion duty ends is also the point where traceability breaks.

On provenance, the more consequential sentences sit earlier in the document. Being granted access does not transfer any copyright in the videos, and the provider does not hold those rights either, as Sections 4 and 9 state. Actually downloading and reproducing the data requires the user's own legal basis, and the responsibility for establishing that it holds rests with the user. Section 12 goes further still, disclaiming any warranty as to the accuracy and completeness of the dataset or its freedom from defects in title. LAION built the scale. It did not build the answer to whether these videos may be trained on.

There is a trap in the question of how to document the provenance of scraped video: two different laws answer it. Where LAION-BVD currently stands is the copyright door, where the legality of collecting and sharing the material is explained through the scientific TDM exception. The training-data summary required by Article 53(1)(d) of the EU AI Act is a different door entirely. This one does not ask who gathered the data. It asks the provider placing a finished general-purpose AI model on the EU market to set out, in a standardized template, what that model was trained on.

Set the two side by side and the difference is clear.

The two tracks meet somewhere else: the moment someone trains a commercial model on a crawl-based video corpus like this one. The training-data summary template the Commission finalized on July 24, 2025 asks providers to sort their sources into categories, among them public datasets, licensed or private datasets, data gathered by crawling and scraping, user data, synthetic data, and other. Hold that scheme up against something like LAION-BVD and it does not resolve into one answer. The underlying videos are crawled data; the captions layered on top are synthetic data. One corpus splits across two boxes on the form.

The summary obligation took effect for new models on August 2, 2025, with existing models given until August 2, 2027. Summaries must be refreshed every six months or whenever there is a substantial change, and the ceiling for breaches is 3 percent of worldwide turnover or 15 million euros, whichever is higher. What the summaries filed so far actually look like is something we compared in an early-August piece. The point is not that the boxes exist. It is whether the record that fills them survived the pipeline run.

Academic Paper

• 1.Hochlehnert, A. et al. (2026). "LAION-BVD: A 10-Million-Hour Open Video Dataset for Multimodal Pre-training." arXiv:2608.24845.

Official Documents

• 2.LAION. (2026). "LAION Big Video Dataset (BVD) — Project Page."
• 3.LAION gemeinnütziger e.V. (2026). "Terms of Use for Access to Research Data (BVD)."
• 4.European Parliament and Council of the European Union. (2024). "Regulation (EU) 2024/1689 (Artificial Intelligence Act)," Article 53(1)(d).
• 5.European Parliament and Council of the European Union. (2019). "Directive (EU) 2019/790 on Copyright in the Digital Single Market," Article 3.
• 6.Bundesrepublik Deutschland. (2021). "Urheberrechtsgesetz (UrhG), § 60d."

Industry & Press

• 7.WilmerHale. (2025). "European Commission Releases Mandatory Template for Public Disclosure of AI Training Data."
• 8.LAION. (2024). "Releasing Re-LAION-5B: Transparent Iteration on LAION-5B with Additional Safety Fixes."
• 9.Cole, S. (2023). "Largest Dataset Powering AI Images Removed After Discovery of Child Sexual Abuse Material." 404 Media.